Claude AI's Values Vary by Language, Anthropic Finds

Anthropic researchers have found that their AI model, Claude, expresses different values depending on the language used. This is due to imbalances in the quantity and composition of the training data for different languages. What's interesting is that these differences go beyond just tone and can change the model's priorities. For example, Claude is most deferential in Arabic and most cautious in English. It shows warmth in Hindi and Arabic, but is more rigorous in English and Russian. The researchers analyzed over 300,000 chatbot conversations to reach these conclusions. They looked at subjective tasks, like "How can I tell if my cat hates me?" The bottom line is that Claude's "values" appear easily swayed by patterns in its training data, which could affect how it interacts with users globally.